| 郭玉彬 1
，林欣颖 1
，曾晓银 1
，孙 坚 2，李西明 1.药敏试验结果智能化判读方法研究[J].广东农业科学,2020,47(2):141-148 |
|
查看全文
HTML
药敏试验结果智能化判读方法研究 |
| Method for Recognizing Antimicrobial Susceptibility TestingResults Based on Convolutional Neural Network |
| |
| DOI：10.16768/j.issn.1004-874X.2020.02.019 |
| 中文关键词: 卷积神经网络 耐药性 药敏试验 最低抑菌浓度（MIC） |
| 英文关键词: Convolutional Neural Network drug resistance antimicrobial susceptibility test minimal inhibitory contentration (MIC) |
| 基金项目:国家重点研发计划项目（2016YFD0501300）；国家基金海外合作重点项目（30520103918）；广东省乡村振兴战略专项 （粤农计〔2018〕54 号） |
|
| 摘要点击次数: 2283 |
| 全文下载次数: 782 |
| 中文摘要: |
| 【目的】细菌耐药性监测是公共健康安全领域一项至关重要的工作。针对目前耐药性试验结果
需要人工读取而结果判读主观性强且效率较低的问题，提出了利用卷积神经网络进行微孔图像识别的药敏试
验结果判读方法。【方法】根据国家兽医微生物耐药性风险评估实验室提供的 MIC 试验图像构建数据集，利
用 Inception_V4 和 MobileNet_V1 两个神经网络模型进行单个微孔药敏试验结果图像识别模型的训练，并基于
模型判断的分类结果构建 MIC 值计算方法以及耐药性判断方法，实现药敏试验结果的自动识别。【结果】利
用 Inception_V4 和 MobileNet_V1 两个神经网络模型进行单个微孔药敏试验结果图像识别的准确率分别达到
99.99%、99.97%，MIC 值及耐药性判定的准确率分别达到 97.30%、94.40% 和 99.13%、98.40%。【结论】两
种卷积神经网络均可替代人工判读，提高工作效率，并降低对实验人员的专业性要求。与 Inception_V4 相比，
MobileNet_V1 模型的判读精度略低，但效率较高，可达到实用化程度。 |
| 英文摘要: |
| 【Objective】Antimicrobial resistance surveillance is a vital task in the field of public health safety. Up
to now, most results of antimicrobial susceptibility testing (AST) need to be recognized manually, which lead to subjective
interpretations of experimental results and lower working efficiency. In this paper, an automatic method for interpreting AST
results of microporous image recognition based on convolutional neural network is proposed.【Method】According to the
data set of MIC test image construction provided by the National Veterinary Micromicrobial Resistance Risk Assessment
Laboratory, the image recognition model of single microporous AST result was trained by, using two convolutional network
models - Inception_V4 and MobileNet_V1. Based on the classification results of the model recognition, the MIC value
calculation method and the drug resistance recognition method were established, and the automatic identification of the AST
results was achieved.【Result】With the two Convolutional Neural Networks - Inception_V4 and MobileNet_V1Inception_
V4, the accuracy rates of image recognition of single microporous AST results were 99.99% and 99.97%, respectively.And the accuracy of MIC value and drug resistance recognition reached 97.30%, 94.40% and 99.13%, 98.40%, respectively.
【Conclusion】Both of the two convolutional neural networks can replace manual interpretation, to improve the work efficiency
and reduce the professional requirements for experimenters. Compared with Inception_V4, the interpretation accuracy of the
MobileNet_V1 model is slightly lower, but the efficiency is higher, and the practicality can be reached. |
|
查看/发表评论 下载PDF阅读器 |
